Microsoft Copilot readiness and secure AI deployments

Microsoft's documentation is clear on the key point: Microsoft 365 Copilot only surfaces organizational data to which individual users have at least view permissions. It doesn't bypass security. It uses the permissions you already have, and that is the problem. Years of 'share with everyone' links, broad Teams membership, inherited SharePoint permissions and old project sites mean a typical user can open thousands of files they have never seen. Copilot makes that content easy to find and summarize.

Microsoft also states that prompts, responses and data accessed through Microsoft Graph are not used to train the foundation large language models, that Copilot honours usage rights on content encrypted by Microsoft Purview Information Protection, and that Copilot is covered by the data residency commitments in its Product Terms. Web search, when enabled, sends generated search queries to the Bing service, which is an admin-controlled setting worth reviewing.

The readiness review answers one question: if every user had Copilot tomorrow, what could they find that they shouldn't? We check:

• SharePoint sites and Teams with organization-wide or public access, and the sensitive content inside them.
• 'Anyone' and organization-wide sharing links, and how many are active.
• Broken permission inheritance and sites with no current owner.
• Large groups (such as 'Everyone except external users') granted access to sensitive libraries.
• OneDrive folders shared broadly by former or current staff.
• Guest accounts and external sharing settings.
• Mailbox delegation and shared mailbox access.
• Licensing and identity basics: MFA, conditional access and admin roles.

Permissions cleanup reduces exposure; labels keep it reduced. We design a small set of Microsoft Purview sensitivity labels (for example Public, Internal, Confidential, Highly Confidential) and apply them where they matter most: HR, finance, legal and client files. Labels that apply encryption restrict who can open content, and Copilot respects those restrictions. Where your licensing supports it, we add data loss prevention policies and auto-labelling for common sensitive information types, such as social insurance numbers or banking details.

We also set retention for Copilot interactions. Microsoft states admins can use Microsoft Purview to set retention policies for data related to chat interactions with Copilot, so prompts and responses can be kept or deleted in line with your records policy.

1. 01

   Pick the pilot group

   Ten to thirty users across departments, including at least one person in a sensitive role, so oversharing is found in a controlled setting.

2. 02

   Set success measures

   Agree what 'useful' means for each team and how you will measure it, so the licence decision is based on evidence.

3. 03

   Train for safe use

   Short sessions on prompting, checking output, and what not to do, aligned with your AI acceptable-use policy.

4. 04

   Red-team the tenant

   Pilot users deliberately search for salary data, HR files, board papers and client records. Anything they find gets fixed before wider rollout.

5. 05

   Decide and expand

   Roll out in waves, with permissions and label checks repeated for each new group.

After rollout, we review Copilot activity in the Microsoft 365 audit log, track sharing and label changes, and recheck oversharing reports on a schedule. New sites and Teams are created constantly; a quarterly check keeps the gains from eroding. The Canadian Centre for Cyber Security's guidance on generative AI (ITSAP.00.041) is a useful reference for staff training, covering risks like inaccurate output, data exposure and AI-assisted phishing.

This work is scoped in writing before work starts. Fees are quoted per engagement. Policy and approval questions about AI in general are covered by our AI governance service; this page is the technical rollout.

• Oversharing and permissions report, ranked by risk
• Remediation of priority sites, links and groups (or a runbook for your team)
• Sensitivity label scheme and configuration
• Copilot settings review, including web search and plugin or agent controls
• Retention configuration for Copilot interactions
• Pilot plan, training material and results summary
• Ongoing review checklist